Self-Organized Neural Integrators in Noisy Spiking Networks

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Neural Dynamics

Background:

• Neural integrators are crucial for cognitive functions like working memory and decision-making.
• Traditional models often require precise recurrent connectivity, which may not be biologically realistic.

Purpose of the Study:

• To explore a biologically plausible mechanism for neural integration in randomly connected spiking networks.
• To understand the role of noise and plasticity in network integration.
• To connect theoretical models to experimental findings in decision-making tasks.

Main Methods:

• Utilizing mean-field theory (MFT) to analyze network dynamics.
• Investigating a local, reward-modulated two-trace plasticity rule.
• Comparing model predictions with experimental data from a tactile decision-making task.

Main Results:

• Randomly connected noisy spiking networks can approximate linear integration under specific parameter conditions.
• Network dynamics are governed by mean recurrent and feedforward weights, with noise being critical.
• The model successfully reproduces adaptive cortical dynamics observed during timing-related learning.
• The framework links to oculomotor persistence and evidence accumulation, modeling drift-diffusion dynamics.

Conclusions:

• Noise-driven, randomly connected networks offer a viable alternative to finely tuned networks for neural integration.
• A specific plasticity rule enables networks to learn and adapt integration properties.
• The unified framework explains diverse neural functions, including working memory, decision-making, and motor control.